Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
1© 2018 arcadia.io | not for redistribution | confidential
CARE MANAGEMENT: USING PREDICTIVE ANALYTICS TO CHOOSE THE
“ACTIONABLE” PATIENTSRICH PARKER, MD
May 16, 2019
2© 2018 arcadia.io | not for redistribution | confidential
INTRODUCTIONCARE MANAGEMENT AND PREDICTIVE ANALYTICS
Rich Parker, MDChief Medical Officer
Arcadia
3© 2018 arcadia.io | not for redistribution | confidential
ABOUT ARCADIAARCADIA OVERVIEW
ARCADIA HAS ANALYZED OVER 50 MILLION PATIENTS NATIONALLY
BOSTON20 Blanchard Rd. #10Burlington, MA
CHICAGO630 E Jefferson St.Rockford, IL
PITTSBURGH965 Greentree Rd Suite 200 Pittsburgh, PA 15220
ARCADIA IS A POPULATION HEALTH MANAGEMENT COMPANY,SPECIALIZING IN DELIVERING VALUE-BASED CARE ANALYTICS AND CAREMANAGEMENT TO THE ENTERPRISE. WE CULTIVATE HIGH QUALITY DATAASSETS ENABLING OUR CUSTOMERS TO EFFECTIVELY SHARE RISK.
SEATTLE1601 5th Ave.Seattle, WA
50M PATIENTSMEASURED
2.5x HIGHER ACOPERFORMANCE
$8.9M AVG SAVINGSIN MSSP
40+ EHR VENDORS
CONNECTED
2002 YEARFOUNDED
No.1 VALUE BASED SERVICES
4© 2018 arcadia.io | not for redistribution | confidential
AGENDACARE MANAGEMENT AND PREDICTIVE ANALYTICS
1
2
3
Why do we need better patient stratification?
How do predictive analytics work?
How are health systems using predictive analytics for stratification?
4 How can predictive analytics be implemented to support care management?
5 Where do we go from here?
5© 2018 arcadia.io | not for redistribution | confidential
WHY DO WE NEED BETTER PATIENT STRATIFICATION?
6© 2018 arcadia.io | not for redistribution | confidential
CASE STUDYCARE MANAGEMENT AND PREDICTIVE ANALYTICS
Manny is a 62 year old man who is a single parent of a teenage son. He is unemployed and on disability. He used to work as a chemist and then a journalist
until he began having seizures that made him unable to work. He had personality changes and run-ins with the law. Due to violent outbursts beyond his control, he
spent time in jail. He does not smoke or drink alcohol.
He is stable now, but has morbid obesity, a seizure disorder, chronic lymphedema and cellulitis of the legs, asthma, diabetes and chronic kidney disease. He is on
multiple medications, including a high dose of valium which helps control his seizures. Many providers don’t understand his need for the valium and try to
discontinue it when I am not present.
He has an excellent relationship with his son who is a junior in High School, and playing on the baseball team.
7© 2018 arcadia.io | not for redistribution | confidential
POPULATION HEALTH CARE MANAGEMENT AND PREDICTIVE ANALYTICS
POPULATION HEALTH -- from the Arcadia perspective -- is the ability to aggregate multiple sources of disparate patient related data, including claims,
EHR data, and medications and use that curated data set for improving the health of the population, and at the same time optimizing the financial
performance of the health care entity. Specific tasks within population health include: identifying and closing gaps in care, leveraging variation in utilization data to improve underutilized care and decrease overutilization, identify and
correct under or overcoding, power care management, disease management and other registry based activities and support sophisticated
financial management.
8© 2018 arcadia.io | not for redistribution | confidential
CARE MANAGEMENT CHALLENGESCARE MANAGEMENT AND PREDICTIVE ANALYTICS
THE RIGHT PATIENTS Who are the right patients to care manage? What is the right volume? Over selecting is wasteful and under selecting undermines value.
PANEL SIZES How many patients should a care manager manage? How intense do you want the interventions to be?
MULTIPLE PROGRAMS Will you also be running disease management programs for heart failure, CKD, COPD, DM, BH, and End Of Life Care? If so, how will they dovetail with your care management programs?
9© 2018 arcadia.io | not for redistribution | confidential
HOW TO DETERMINE ROI ON CARE MANAGEMENTCARE MANAGEMENT AND PREDICTIVE ANALYTICS
Faith-based method: “We know it works, and we don’t have to expend resources proving it.”Retrospective review of cohort utilizationCase control method
10© 2018 arcadia.io | not for redistribution | confidential
CARE MANAGEMENT AND PREDICTIVE ANALYTICS
IMPACTABILITY: Assign patients to nurse care manager panels based on their likelihood of benefiting from care management.
HIGH COST / HIGH RISK
RISING COST / RISING RISK
LOW COST / LOW RISK
5-10 nurse care managers…
…serving 100,000 patients.
RISK ≠ IMPACTABILITY
CARE MANAGER EMPANELMENT
Which patients could benefit from care management?
11© 2018 arcadia.io | not for redistribution | confidential
WHAT DOES CARE MANAGEMENT SUCCESS LOOK LIKE?CARE MANAGEMENT AND PREDICTIVE ANALYTICS
Which patients are most likely to respond to care management? How do we measure that?
improvements in condition reductions in cost and utilization
12© 2018 arcadia.io | not for redistribution | confidential
CHALLENGES WITH TRADITIONAL STRATIFICATION APPROACHESCARE MANAGEMENT AND PREDICTIVE ANALYTICS
Highest risk/cost patients are not generally impactable with care management (cancer, accidents)
Traditional risk algorithms are designed for risk adjustment more than population stratification
Traditional risk algorithms do not include all the data needed to predict who will benefit from care management
13© 2018 arcadia.io | not for redistribution | confidential
WHAT IS PREDICTIVE ANALYTICS?CARE MANAGEMENT AND PREDICTIVE ANALYTICS
It’s not about finding out things that are happening right now.
It’s not about finding out exact outcomes in the future.
It is about using existing information to identify patterns and to infer trends and potential outcomes in the future.
“How often are my diabetics going to the ED?”
“Which diabetics are going to end up in the ED next year?”
“Which diabetics are likely to use the ED – but could be steered elsewhere?”
14© 2018 arcadia.io | not for redistribution | confidential
WHAT CAPABILITIES DO WE NEED?CARE MANAGEMENT AND PREDICTIVE ANALYTICS
Requirements for a better approach to stratificationDoes not just identify the sickest or highest cost patients.Can be used in a variety of contexts and populations.Can be used to report on diverse individuals regardless of background.Can help clinicians identify clusters of patients within a population for inclusion in programs
15© 2018 arcadia.io | not for redistribution | confidential
HOW DOES IT WORK?THE ARCADIA IMPACT SCORE IS A PREDICTIVE MODEL DEVELOPED FROM A MACHINE LEARNING ALGORITHM BASED ON AGGREGATED DATA
16© 2018 arcadia.io | not for redistribution | confidential
FINDING THE IMPACTABLE COHORTHOW DOES IT WORK?
$ PMPM 17% -18% -26%
% Leakage 20% -17% -21%
Hospitalizations -23% -25% -36%
Avoidable Admits 22% -24% -30%
ED Visits 13% -37% -40%
Avoidable ED Visits 11% -35% -43%
Medications 9% 6% 11%
17© 2018 arcadia.io | not for redistribution | confidential
BUILDING AN IMPACTABILITY MODELHOW DOES IT WORK?
EXPECTED CM VALUE:▪ Cost▪ Utilization▪ Health Outcomes
A
B
UNMANAGEDPOPULATION
ENROLLED& MANAGED
INPUTS▪ Demographics▪ Morbidity Risk▪ Condition Types▪ Utilization (OP/IP)▪ Census Factors▪ Care Coordination▪ Population Flags
OUTPUTS▪ Cost▪ Utilization▪ Outcomes
18© 2018 arcadia.io | not for redistribution | confidential
POTENTIAL MODEL INPUTSHOW DOES IT WORK?
COST STRATA▪ 12-month true cost▪ 6-month true cost▪ Predicted future cost
SOCIAL STRATA▪ Education levels▪ Income levels▪ Public assistance▪ Access to transportation▪ Veteran, Homeless
indicators▪ Home status
MORBIDITY STRATA▪ 100 disease groups▪ Concurrent morbidity risk▪ Dx clusters▪ Expected mortality
UTILIZATION STRATA▪ Recent ED utilization▪ Recent IP utilization▪ Medication utilization and
polypharmacy▪ IP and ED predicted
utilization
CARE STRATA▪ Care coordination risk▪ Care team density▪ Frailty▪ Hospice
PERFORMANCE STRATA▪ Quality gaps▪ Activation and adherence▪ Coding gaps▪ Coding/Quality and
revenue opportunity▪ Wellness visit adherence
COMBINED EHR/CLAIMS/ADMIN LONGITUDINAL RECORD
PREDICTIVE IMPACTIBILITY SCORES
19© 2018 arcadia.io | not for redistribution | confidential
CENSUS DATAHOW DOES IT WORK?
Where SDOH data are not provided on a patient level, census data from the American Community Survey presented at a Census Block Group level can provide a high-resolution picture of socioeconomic status.
Census Block Group in Cambridge,
MA
Population% Males% Females% Under 18% 18 - 44% 45 - 64% 65+% High School% Bachelors% Graduate DegreeMedian Earnings (Real Dollars)Female Earning Ratio (Median Female Earnings/Median Male Earnings)% Population by Race% Population by Race - American Indian or Alaskan Native% Population by Race - Asian
% Population by Race - Hispanic% Population by Race - Black% Population by Race - White Non-HispanicPersons per Housing Unit% Families w/ Incomes < 100% of Federal Poverty Level% Families w/ Incomes < 200% of Federal Poverty Level% Adults who are Unemployed% Households Receiving Public Assistance% Households w/ No Car% Households with Children and a Single Parent% People Age 25+ w/o High School Degree
20© 2018 arcadia.io | not for redistribution | confidential
FACTOR SELECTIONHOW DOES IT WORK?
Of all the input factors supplied to the model, most can be eliminated due to their limited
effects on the model, in this case using a process called Backwards Factor Selection
21© 2018 arcadia.io | not for redistribution | confidential
BUILDING AN IMPACTABILITY MODELHOW DOES IT WORK?
EXPECTED CM VALUE:▪ Cost▪ Utilization▪ Health Outcomes
A
B
UNMANAGEDPOPULATION
ENROLLED& MANAGED
INPUTS▪ Demographics▪ Morbidity Risk▪ Condition Types▪ Utilization (OP/IP)▪ Census Factors▪ Care Coordination▪ Population Flags
OUTPUTS▪ Cost▪ Utilization▪ Outcomes
A
B
22© 2018 arcadia.io | not for redistribution | confidential
FACTOR SENSITIVITIESHOW DOES IT WORK?
23© 2018 arcadia.io | not for redistribution | confidential
IMPACTSCORE
BUILDING AN IMPACTABILITY MODELHOW DOES IT WORK?
EXPECTED CM VALUE:▪ Cost▪ Utilization▪ Health Outcomes
A
B
UNMANAGEDPOPULATION
ENROLLED& MANAGED
INPUTS▪ Demographics▪ Morbidity Risk▪ Condition Types▪ Utilization (OP/IP)▪ Census Factors▪ Care Coordination▪ Population Flags
OUTPUTS▪ Cost▪ Utilization▪ Outcomes PROJECTED IMPACT
OF CARE MGMT
PROJECTED IMPACT OF CARE MGMT
IMPACTSCORE
The impact score describes the relative benefit projected for this patient from care management.
A
B
24© 2018 arcadia.io | not for redistribution | confidential
HOW IS IT USED?FLEXIBLY! APPLICATION OF THE IMPACT SCORE DEPENDS ON THE NEEDS OF THE HEALTHCARE ORGANIZATION
25© 2018 arcadia.io | not for redistribution | confidential
APPLICATIONSHOW IS IT USED?
26© 2018 arcadia.io | not for redistribution | confidential
APPLICATIONSHOW IS IT USED?
There are a number of ways the impact score can be used…Rank StratificationScore Clustering
Population-Driven SegmentationInitiative-Driven Segmentation
• Classification
27© 2018 arcadia.io | not for redistribution | confidential
APPLICATION: FIND CANDIDATES FOR CARE MANAGEMENTHOW IS IT USED?
Rank Segmentation
How: Rank your population by Impact Score and transmit the top “N” individuals for assessment.Why: Simple way to sort the population and allow grass-roots professionals to make decisions.Caveat: Assumes all CMs are the same, no consideration of circumstances or conditions.
We want to give our Care Managers lists of likely
candidates to review for entry into a program.
28© 2018 arcadia.io | not for redistribution | confidential
APPLICATIONS: LOAD CARE MANAGEMENT QUEUESHOW IS IT USED?
Score Clustering
How: Group individuals into clusters by score (high/medium/low) and choose groups for assessment.Why: Simple way to sort the population, but give grass-roots professionals more options.Caveat: Requires more filtering by CMs, and still doesn’t consider circumstances or conditions.
At any given time, we want to have a target number of
patients enrolled in our care management programs.
29© 2018 arcadia.io | not for redistribution | confidential
APPLICATION: FIND PATIENTS WITH SPECIFIC ATTRIBUTESHOW IS IT USED?
Population-Driven Segmentation
How: Segment your population by specific attributes, and then rank by score and transmit to CMs.Why: Deliberate focus on features of the population drives decisions on care management.Caveat: More complex process of determining segmentation; could ignore critical corner cases.
We want to find the most actionable patients within
certain segments of our population.
30© 2018 arcadia.io | not for redistribution | confidential
APPLICATION: QUALIFY PATIENTS FOR PROGRAMSHOW IS IT USED?
Initiative-Driven Segmentation
How: Group individuals by qualification into specific initiatives, and then rank and transmit.Why: Optimizes guidance of the score against actual use cases, making results more meaningful.Caveat: Requires thoughtful initiative design and consideration of criteria; plus, not all initiatives apply to the same score.
We have planned a diabetes management initiative and
want to find the right candidates.
31© 2018 arcadia.io | not for redistribution | confidential
APPLICATIONSHOW DO WE IMPLEMENT IT?
Predictive analytics offer insights that can be powerful, but that can also be counterintuitive.
Therefore, scores such as the Arcadia Impact Score should be just one tool among many in your risk stratification toolbox.
32© 2018 arcadia.io | not for redistribution | confidential
WHERE DO YOU GO NEXT?PREDICTIVE TOOLS CAN HELP YOU BETTER UNDERSTAND OPPORTUNITIES AT A POPULATION LEVEL
33© 2018 arcadia.io | not for redistribution | confidential
QUESTIONS?WHERE NEXT?
[email protected]://www.linkedin.com/in/rich-parker-b998a81a/